AN APPLICATION OF BRANCH AND CUT TO OPEN PIT MINE

SCHEDULING
Louis Caccetta and Stephen P. Hill
School of Mathematics and Statistics
Curtin University of Technology
GPO Box U1987
PERTH, Western Australia, 6845
caccetta@cs.curtin.edu.au
hillsp@cs.curtin.edu.au
ABSTRACT
The economic viability of the modern day mine is highly dependent upon careful
planning and management. Declining trends in average ore grades, increasing mining
costs and environmental considerations will ensure that this situation will remain in the
foreseeable future. The operation and management of a large open pit mine having a
life of several years is an enormous and complex task. Though a number of
optimization techniques have been successfully applied to resolve some important
problems, the problem of determining an optimal production schedule over the life of
the deposit is still very much unresolved. In this paper we will critically examine the
techniques that are being used in the mining industry for production scheduling
indicating their limitations. In addition, we present a mixed integer linear programming
1
model for the scheduling problems along with a Branch and Cut solution strategy.
Computational results for practical sized problems are discussed.
KEYWORDS: Branch and Cut, Mixed Integer Linear Programming, Mine
Scheduling, Optimization.
1. INTRODUCTION
The operation and management of a large open pit mine is an enormous and
complex task, particularly for mines having a life of many years. Optimization
techniques can be successfully applied to resolve a number of important problems that
arise in the planning and management of a mine. These applications include: ore-body
modelling and ore reserve estimation; the design of optimum pits; the determination of
optimal production schedules; the determination of optimal operating layouts; the
determination of optimal blends; the determination of equipment maintenance and
replacement policies; and many more (Caccetta and Giannini [7-9]).
A fundamental problem in mine planning is that of determining the optimum
ultimate pit limit of a mine. The optimum ultimate pit of a mine is defined to be that
contour which is the result of extracting the volume of material which provides the total
maximum profit whilst satisfying the operational requirement of safe wall slopes. The
ultimate pit limit gives the shape of the mine at the end of its life. Usually this contour
is smoothed to produce the final pit outline.
Optimum pit design plays a major role in all stages of the life of an open pit: at
the feasibility study stage when there is a need to produce a whole-of-life pit design; at
the operating phase when pits need to be developed to respond to changes in metal
prices, costs, ore reserves, and wall slopes; and towards the end of a mines life where
the final pit design may allow the economic termination of a project. At all stages there
2
is a need for constant monitoring of the optimum pit, to facilitate the best long-term,
medium-term and short-term mine planning and subsequent exploitation of the reserve.
The optimum pit and mine planning are dynamic concepts requiring constant review.
Thus the pit optimization technique should be regarded as a powerful and necessary
management tool. Further, the pit optimization method must be highly efficient to
allow for an effective sensitivity analysis. In practice one needs to construct a whole
spectrum of pits, each corresponding to a specific set of parameters.
The ultimate pit limit problem has been efficiently solved using the Lerchs-
Grossmann [28] graph theoretic algorithm or Picards [33] network flow method (see
also Caccetta and Giannini [7,8]). These methods are based on the block model of an
orebody; the block model is detailed in the next section. A comparative analysis of the
two methods is given by Caccetta et al [10]. Optimum pit design plays an important
role in mine scheduling.
The open pit mine production scheduling problem can be defined as specifying
the sequence in which blocks should be removed from the mine in order to maximise
the total discounted profit from the mine subject to a variety of physical and economic
constraints. Typically, the constraints relate to: the mining extraction sequence;
mining, milling and refining capacities; grades of mill feed and concentrates; and
various operational requirements such as minimum pit bottom width. The scheduling
problem can be formulated as a mixed integer linear program (MILP). However, in real
applications this formulation is too large, in terms of both the number of variables and
the number of constraints, to solve by any available MILP software.
Several approaches to the scheduling problem have appeared in the literature
including: heuristics (Caccetta et al [14] and Gershon [23]); Lagrangian relaxation
(Caccetta et al [14]); parametric methods (Dagdelen and Johnson [17], Francois-
3
Bongarcon and Guibal [6], Matheron [29,30] and Whittle [37,38]); dynamic
programming techniques (Tolwinski and Underwood [36]); mixed integer linear
programming (Caccetta et al [11,14], Dagdelen and Johnson [17] and Gershon [22]);
and the application of artificial intelligence algorithms such as simulated annealing
genetic algorithms (Denby and Schofield [18]) and neural networks (Denby et al [19]).
Because of the complexity and size of the problem all these approaches suffer from one
or more of the following limitations: cannot cater for most of the constraints that arise;
yield only suboptimal solutions and in most cases without a quality measure; can only
handle small sized problems.
In this paper we will present the result of our effort to produce a computational
method that incorporates all the constraints in the optimization and yields provably
good solutions for reasonably large size problems. We give a MILP formulation of the
scheduling problem and present a Branch and Cut procedure for its solution. This is
done in Section 3. Computations results are discussed in Section 4. The next section
provides details of the block model used as well as a critical account of the various
techniques that are used in the mining industry. We only discuss the more promising
methods that have been applied to real mines.
2. PRELIMINARIES
In this section we outline some of the methods that have been proposed for
various mine development problems. We begin with the basic block model of an ore
body, then we present a mixed integer linear programming formulation of the
scheduling problem and discuss a number of algorithms that have been proposed for its
solution. We focus mainly on methods which have proved useful in the mining
industry. For a discussion of other methods we refer to Kim [27] and Thomas [34,35].
4
2.1 Block Model
An early task in mine management is the establishment of an accurate model for
the deposit. Though a number of models are available, the regular 3D fixed-block
model is the most commonly used and is the best suited to the application of
computerized optimization techniques (Gignac [24] and Kim [27]). This model is
based on the ore body being divided into fixed-size blocks. The block dimensions are
dependent on the physical characteristics of the mine, such as pit slopes, dip of deposit
and grade variability as well as the equipment used. The centre of each block is
assigned, based on drill hole data and a numerical technique, a grade representation of
the whole block. The numerical technique used is some grade extension method such
as : distance weighted interpolations, repression analysis, weighted moving averages
and kriging (Gignac [24]). Using the financial and metallurgical data the net profit of
each block is determined.
The wall slope requirements for each block are described by a set (typically 4
to 8) of azimuth-dip pairs. From these we can identify for each block x the set of
blocks S
x
which must be removed before block x can be mined. This collection of
blocks, x S
x
, is usually referred to as a cone .
The key assumptions in the block model are : the cost of mining each block
does not depend on the sequence of mining; and the desired wall slopes and pit shape
can be approximated by the removed blocks.
2.2 The Scheduling Problem
The open pit mine production scheduling problem can be defined as specifying
the sequence in which blocks should be removed from the mine in order to maximize
5
the total discounted profit from the mine subject to a variety of constraints. The
constraints may involve the following :
mill throughput (mill feed and mill capacity)
volume of material extracted per period
blending constraints
stockpile related constraints
logistic constraints
We now present a simple mixed integer linear programming (MILP) formulation that
incorporates the mill throughput and volume of material extracted constraints. We
begin with some notation. Let
T is the number of periods over which the mine is being scheduled.
N is the total number of blocks in the ore body.
t
i
c is the profit (in NPV sense) resulting from the mining of block i
in period t.
O is the set of ore blocks.
W is the set of waste blocks.
t
i
is the tonnage of block i.
m
t
is the tonnage of ore milled in period t.
S
i
the set of blocks that must be removed prior to the mining of
block i.

'

otherwise. 0,
t to l periods in mined is i block if 1,
t
i
x
t
0
l lower bound on the amount of ore that is milled in period t.
t
0
u upper bound on the amount of ore that is milled in period t.
6
t
w
u upper bound on the amount of waste that is mined in period t.
Then the MILP formulation is :
Maximize Z =
T
i
N
1 i
T
i
1 t
i
t
i
1 t
i
N
1 i
T
2 t
x c x ) c c (

+
(2.1)
subject to
1
1
i
O i
i
m x t

(2.2)
t 1 t
i
t
i
O i
i
m ) x x ( t

, t=2,3,...,T. (2.3)
1
w
1
i
W i
i
u x t

(2.4)
, u ) x x ( t
t
w
1 t
i
t
i
W i
i

t = 2,3,...,T. (2.5)
t
i
1 t
i
x x

, t = 2,3,...,T. (2.6)
t
j
t
i
x x
, t=1,2,...,T, j S
i
; i = 1,2,...,N. (2.7)
t
0
t t
0
u m l
, t = 1,2,...,T. (2.8)
1 , 0 x
t
i
, for all i, t. (2.9)
Constraints (2.2), (2.3) and (2.8) ensure that the milling capacities hold. Constraints
(2.4) and (2.5) ensure that the tonnage of waste removed does not exceed the prescribed
upper bounds. Constraints (2.6) ensures that a block is removed in one period only.
Constraints (2.7) are the wall slope restrictions.
The above formulation has NT 0-1 variables, and (N+2)T + N(d-1) linear
constraints, where d is the average number of elements in a cone. Typically T is around
10, N is 100,000 for a small pit and over 1,000,000 for a larger pit. Consequently the
MILPs that arise are much too large for direct application of commercial packages.
However, as we demonstrate in this paper, the structure of the problem can be exploited
to develop computational strategies that produce provably good solutions.
7
Solving MILPs such as (2.1) (2.9) is a difficult and challenging task. Indeed,
in the mining context, the lack of an immediate optimization technique has led the
mining industry to focus on easy subproblems. Consequently the schedules that are
generated, usually manually, are often outside the specified operating range and
certainly far from optimal. The usual approach is to first determine the final pit outline
and then through a series of refinements mining schedules are generated.
The final pit outline is determined by smoothing the contour produced by
solving the ultimate pit limit problem. The ultimate pit limit is the maximum value pit
resulting from the mining of ore and waste blocks under the assumption that all mining
could be done in one period. That is, the solution to the problem (2.1) subject to (2.7)
with T = 1 and (2.9).
The ultimate pit limit problem can be solved using the Lerchs-Grossmann graph
theoretic algorithm [28] or by solving Picards [33] network flow formulation. Over
the past 10 to 15 years efficient packages have become available for solving this
problem (Caccetta et al. [10,11], Whittle [37,38] ). Prior to this, the Moving Cone
Technique was widely used because of its simplicity. This technique basically selects a
block x for mining providing the total profit from blocks contained in the cone x S
x
is
positive. Whilst the method is extremely simple, it is easy to show that solutions far
from optimal can be obtained. A number of refinements to the technique have been
proposed (Yamatomi et al. [40]).
The Lerchs-Grossmann Algorithm (LGA) provided an important tool for mine
design. However, as time is not an input parameter, its use for scheduling is restricted
to mines having a very short life (up to 3 years). In the following we detail various
approaches to mine scheduling.
2.3 Algorithms
8
We begin by showing that the strategy outlined above for obtaining the
production schedule of a mine by first determining the final pit outline and then
generating the schedule is sound.
Let
C
u
: contour generated by the application of the Lerchs-Grossman
algorithm (LGA).
C
s
: final contour generated by an optimal schedule.
We note that the LGA generates a contour C
u
with minimal number of blocks. We
assume that the contour C
s
also has a minimal number of blocks. Note that the C
s
contains all blocks mined over the T time periods, that is :
C
s
=
i
T
1 i
i
P : P {

set of blocks mined in period i}.

Theorem : Consider an open pit mine in which all constraints have a non-negative
upper bound and a zero lower bound. Then C
s
C
u
.
Proof : Suppose to the contrary that C
s

/ C
u
. Let C
1
= C
s
\C
u
be the set of blocks
mined under the optimal schedule that do not lie in C
u
. Consider
C
2
= C
s
\C
1
= C
s
C
u
,
the set of blocks mined under the optimal schedule that lie in C
u
. Observe that C
2
is a
feasible pit. Further,
Z(C
s
) = Z(C
1
) + Z(C
2
),
where Z(C
i
) is the total value of the blocks in contour C
i
, i = 1,2. The minimality of C
s
implies that Z(C
i
) > 0 for each i. Now consider the contour
C = C
u
C
1
.
9
Observe that C satisfies the wall slope restrictions and consequently is a feasible
contour in terms of the ultimate pit limit problem. Further, since C
u
C
1
= ,
contour C has a total value of
Z(C ) = Z(C
u
) + Z(C
1
)
> Z(C
u
),
a contradiction. This proves that C
s
C
u
.
As mentioned in the previous subsection, good commercial packages are
available for obtaining the contour C
u
. Below we detail approaches that attempt to
determine the production schedule.
Parameterization Method
In their paper Lerchs and Grossmann [28] introduced the concept of parametric
analysis in order to generate an extraction sequence. They considered the undiscounted
model and varied the economic value of each block i from c
i
to (c
i
- ) for varying
0. An increasing sequence of values gives rise to a set of nested pits. These pits can
be used to produce a production schedule. Since this early work a number of authors
have considered the implementation aspects of this method and its variations (Francois-
Bongarcon and Guibal [6], Caccetta et al. [11,14], Calou [15], Dagdelen and Johnson
[17], Matheron [29,30], and Whittle [37,38]).
The mostly widely used scheduling software package that is based on
parameterization, is Whittles Four-D and Four-X [37,38]; the latter allows for multiple
ore types in the calculation of block costs. The parameter used in theses packages is
referred to as the metal cost of mining which is defined as : extraction cost
($/ton)/selling price ($/gm). This quantity provides an indication of the amount of
product that must be sold to cover the cost of extracting one ton of material. The
10
rationale for using this parameter is that the three components for calculating the block
values (selling price ($/unit), processing cost ($/ton) and extraction cost ($/ton)) reduce
to one factor under the assumption that the ratio of processing and extraction costs is
constant. Whittle [37,38] suggests that a best mining schedule comes from extracting
each of the nested pits in turn and a worst schedule comes from extracting the ore
bench by bench.
The advantages of the Whittle approach include :
the nested pits can be determined efficiently as each requires the
solution of an ultimate pit limit problem.
the identification of clusters of high grade ore in the model.
a measure in the design of the final pit contour subject to a change in
price and thus some sensitivity analysis can be performed.
The disadvantages of the Whittle approach includes :
time and other variable factors (for example, extraction rate, different
ore types, blending, etc.) are only implicitly included in the optimization
through modifying the cost function.
the possibility of a large increment in the size of the pit from one nested
pit to the next. This is referred to as the gapping problem and it arises
because there is no clear method for choosing the values of .
optimality is not guaranteed. Indeed the best schedule may not even
provide an upper bound for the NPV of the mine. This is easily seen by
noting the nested pits produced by the LGA cannot have waste at the bottom
whereas those produced by an optimal schedule can.
11
the validity of the underlying assumption that the rate of processing
and extractor costs are constant. Indeed, Whittle [37] states that variation of
t 20% over 5 years is typical.
the necessity of reducing time costs to a cost per ton basis. Making
assumptions about the production rates, these costs are determined
iteratively until a reasonable solution is found.
structural constrains such as mining to a maximum vertical depth or a
minimum mining width, cannot be incorporated into the parameterization
model.
the extraction sequence may not satisfy the production requirements of
the mine.
Recently, in an attempt to address the spacial requirements, Whittle [38] introduced the
Milawa Algorithm which given a set of nested pits produces a revised schedule with the
aim of improving the NPV.
Another commercially available package which extends the nested pit approach
is the Earthworks NPV Scheduler [20]. This package first generates the nested pits and
then using these, the pushbacks are defined heuristically. The criteria for the pushbacks
is to keep them as close as possible to the extraction sequence suggested by the nested
pits taking into consideration equipment access. Finally, a restricted tree search
procedure is used to resequence the pushback removal to increase the NPV. A major
advantage of this package is that it may produce schedules that are more likely to be
acceptable to mining engineers because practical spacial constraints are taken into
acccount when defining the pushbacks.
12
Unfortunately, all methods that use the above nested pits approach in a
sequential optimization procedure may produce a schedule that varies considerably
from the optimum. Indeed, even a feasible solution cannot be guaranteed.
MILP Approach :
A number of authors have proposed MILP formulations to various mine
scheduling problems (Caccetta et al. [11,14], Dagdelen and Johnson [17], Gershon
[21,22] and Kim [27]). The major computational difficulty has been the size of the
problem. Typically MILP approaches are developed in house for short term
schedules. Below we outline two approaches for solving the MILP formulations.
Recently, Combinatorics Pty Ltd has released the package MineMax [31] for
long term mine scheduling. The MILP is solved using a commercial package (for
example, CPLEX). Our understanding is that if the problem is too large for the MILP
solver, or if a solution is not obtained within a prescribed time period, then a second
option is offered. This option is to solve each MILP formulation for free variables on a
period by period basis.
The advantages of MineMax include :
all periods are addressed concurrently in the global case.
the optimization incorporates all constraints including grade; structural
considerations; etc. Thus even if larger block sizes are used (to reduce the
number of variables) the solution obtained may be better than that obtained
using the nested pit approach.
The disadvantages of MineMax include :
capable of solving only very small size problems due to the large
number of integer variables and constraints. This is true for both options as
will be illustrated in the computational section.
13
since reblocking is often required, the wall slope requirements are
poorly approximated as is the block data.
in the period by period option, the schedule obtained may be far from
optimum.
Caccetta et al [14] proposed a Lagrangian relaxation method for solving the MILP. At
each step a problem similar to the ultimate pit limit problem is solved using the LGA
with additional constraints dualized. Subgradient optimization is used to reduce the
duality gaps. The method is tested on a real ore body with 20,979 blocks and 6 time
periods. The schedules obtained are within 5% of the theoretical optimum. The main
problem with the method is resolving the duality gaps. However, the subproblems are
useful in producing solutions using a heuristic. In fact, the heuristic solution obtained
for the real ore body is within 2% of the theoretical optimum.
Other Methods :
We conclude this section by briefly mentioning two other approaches. Runge
Mining Pty Ltd have developed the XPAC Autoscheduler package for mine scheduling
[5]. Their heuristic approach is based on the method proposed by Gershon [23] which
iteratively selects blocks to be extracted on a period by period basis. A weighted
function is used to determine the removal sequence. At each step only blocks whose
predecessors have been mined are considered. The advantage of the method is its
speed. Its main use is an interactive tool where the user can see a large number of
scenarios by fixing in and out blocks and running the heuristic. The main
disadvantages are : the search is myopic; no guarantee of finding a feasible solution;
the obtained solution may be far from optimal. The method has been applied to models
with up to 100,000 blocks.
14
Tolwinski and Underwood [36] proposed a method which combines concepts
from stochastic optimization and artificial neural networks with heuristics exploiting
the structure of the mine. The method works by modelling the development of the mine
as a sequence of pits (states) where each pit differs from the previous pit by the removal
of one block (state change). A probability distribution based on the frequency with
which particular states occur is used to determine the state changes. Heuristic rules are
incorporated to learn these characteristics of the sequence of pits which produce a good,
or poor, result.
The main advantages of the method are :
effects of time and other factors can be explicitly included in the
optimization.
the structural constraints are incorporated.
applicable to mines of realistic size (tested on models of up to 88,000
blocks).
The main disadvantages are :
only a small number of all possible sequences can be explored. The
method suffers from combinatorial explosion of the number of states.
no guarantee of finding a feasible solution if one exists.
no measure of the quality of the solution.
3. A NEW BRANCH AND CUT METHOD
In this section we outline our Branch and Cut procedure for solving the MILP
(2.1) - (2.9). Our work is motivated by the recent success of this approach to various
large combinatorial optimization problems including : the travelling salesman problem
(Applegate et al. [3] and Padberg and Rinaldi 32]); the vehicle routing problem
15
(Achuthan et al. [1,2] and Augerat et al. [4]); airline scheduling (Hoffman and Padberg
[25]); and various constrained spanning tree problems (Caccetta and Hill [12,13]).
The objective of our work is to produce a method which explicitly incorporates
all constraints in the optimization and is capable of producing provably good solutions
for reasonably large problems. The quality bound is important as it provides mining
engineers confidence with the results produced.
We now detail some of the important features of our method which exploits the
structure of the problem. Our algorithm has been implemented in C++ and involves
some 17,000 lines of code. The code has been tested on operating mines. Because of
the commericalization of the software and confidentiality agreements we are unable to
provide full details of all aspects of our work. However, we do summarize below some
important features.
Key Features :
1. The block model is reduced to only include blocks inside the final pit design
developed from the ultimate pit (Note: Theorem in previous section).
Further reductions are made through consideration of (2.2) - (2.5) and (2.7).
2. The MILP has strong branching variables due to the dependencies between
variables ((2.6) and (2.7)). Note that setting a variable to 0 or 1 will fix a
potentially large number of other variables. Consequently the subsequent LP
relaxations are significantly smaller in size. This motivates more branching
compared to typical Branch and Cut methods.
3. Cutting planes involving Knapsack constraints are identified using the
capacity upper bounds ((2.2) - (2.5)) and the block removal dependencies
(2.7). Also cuts are identified through material removal dependencies
between benches.
4. Our search strategy involves a combination of best first search and depth first
search. The motivation for this is to achieve a good spread of possible pit
16
schedules (best first search) whilst benefiting from using depth first search
where successive LPs are closely related from one child node to the next.
For large problems this often results in provably good solutions being found
earlier than a search method geared to establishing an optimal solution.
5. Good lower bounds are generated through the use of an LP-Heuristic. The
method works by considering each period in turn and fixing in and out sets
of free variables. Cutting planes are then generated for the period, further
LPs are solved and further fixing occurs. Throughout the fixing of variables
feasibility checks are used. If the heuristic succeeds, or fails due to an
inferior lower bound being found, then periods are considered in the same
direction, otherwise the direction is reversed. The heuristic is called for the
first five levels of the search tree and every eighth node created thereafter.
6. Standard fixing of non basic variables using reduced costs is carried out.
Because of the block dependencies this may lead to the LP solution losing its
optimality. In this case we call the LP solver and re-enter the cutting plane
generation phase without branching.
7. Many branching rules were tested and the following proved to be the best.
The free variables are considered and a subset of these is chosen on the basis
of closeness to the value of 0.5. For each variable in the subset we calculate
the sum change in the fractional values of all variables dependent on the
inclusion and exclusion of the branching variable. Choose the one with the
highest minimal sum change in both directions of branching. Strong
branching is used if the gap between the lower and upper bound is
sufficiently small.
8. If the LP subproblem is not solved within a prescribed maximum time (2
minutes), then the LP optimization is terminated and branching is performed
using the rules in 7 above. An attempt is then made to solve the resulting
17
LPs within the specified time. This process is repeated as long as necessary.
9. The cutting plane phase is terminated early if tailing-off is detected or if the
LP subproblem is solved optimally in more than a prescribed time (1
minute). Note that adding further cutting planes, even with purging of
ineffective constraints, tends to increase the solution times for successive
calls to the LP solver.
10. When branching we probe a random subset of variables having the same
time index as the branching one. Bounds on variables may also be updated
during this process.
11. All our LP subproblems are solved using CPLEX Version 6.0 [16].
Before discussing computational results in the next section we conclude this
section with brief discussion on some extensions to the MILP formulation (2.1) - (2.9)
to incorporate further practical mining restrictions.
Model Extensions:
Besides including constraints from the linear sum of attributes of blocks or the
ratio of quantities and bounding them (for example, blending), the following are
modelled :
Processing different ore types
In many applications we have K 2 ore types which need to be processed through the
mill. The processing rate, in tons/hour, r
i
for ore type i is known. The total processing
time per period constraints can be written as :
, Q x t
r
1
N
1 i
1
i
ik
K
1 k
k

,
_

18
, Q ) x x ( t
r
1
N
1 i
1 t
i
t
i
ik
K
1 k
k

,
_

t = 2,3,,T
where
t
ik
: tonnage of ore type k in block i
Q : expected time (in hours) the mill is available per period.
Maximum vertical depth
To allow pit access (haul roads) it is desirable to restrict the maximum vertical
depth D that can be mined in any one period. This restriction can be modelled as
follows. Suppose blocks i and j have coordinates (x,y,z
1
) and (x,y,z
2
), respectively
with z
1
< z
2
and z
2
- z
1
> D. Then blocks i and j must be mined in different time
periods and so
1
x
l
= 0, for all blocks l more than D units below the surface.
1 t
i
t
j
x x

, for t = 2,3,,T.
Minimum pit bottom width
In order to facilitate equipment movement, a minimum pit bottom width for
each period needs to be specified. We do not believe that linear constraints can be used
to specify this. The usual approach is to manually smooth the base of each
incremental pit; optimality is lost by this process. A feature we have noted of mining is
that when a pit is developed with blocks extracted from several benches the wall slopes
formed are rarely at their upper bounds, except for blocks located at the limits of the
final pit. Consequently, our approach is to redefine the wall slope angles proportionally
19
at all blocks except those located at the limits. This will have the effect that blocks are
removed in clusters and pushbacks can be naturally defined.
Stockpiles
Stockpiles are formed for a number of reasons including: blending; storage of
excessive production; and storage of low grade ore for possible future processing.
When placing an ore block on a stockpile the block characteristics (grade, tonnage, etc)
are known. However, as blocks are mixed on the stockpile, the characteristics of
materials removed from the stockpile to the mill need to be treated as variables. Since
the amount of ore removed from the stockpile is unknown prior to the optimization this
gives rise to some non-linear constraints. To overcome this, we define a variable for
the grams of metal taken from the stockpile per time period as well as tonnage of
material. Then, using these variables, the average grade of ore being removed from the
stockpile is implied and can be rounded. Note that this formulation defines a valid
upper bound for the problem. We use several different constraints to bound the average
grade value as well as conservation of movement constraints to make the upper bound
as tight as possible.
20
4. COMPUTATIONAL RESULTS
Our Branch and Cut algorithm has been implemented in C++ on an SGI Origin
200 dual processor computer. The dual processor was only used to solve the relaxed
LPs. The software has been extensively tested both on test data provided to us by our
industry partner as well as data from producing mines.
The models in our test data range from 26,208 to 209,664 blocks. In all cases T
= 10. We ranged the constraints on the amount of material removed as well as the
bounds on the milling requirements so as to cover the large number of cases that can
actually occur. For the smaller models solutions guaranteed to be within 0.4% of the
optimum were obtained within 12 minutes. For the largest model, solutions guaranteed
to be within 2.5% of the optimum were obtained within 4 hours. For these larger
models we continue the computations for a further 16 hours and observed there was
negligible change in the gap. Our method generates tight bounds. However,
establishing optimality (except, of course, for small problems) is difficult because once
we achieve a near optimal solution there are no available cutting planes to remove
fractional variables occurring in the same bench level. Note that (2.6) and (2.7) give
dependencies between variables corresponding to block removal in time and the vertical
dimension, but not horizontally.
Following the above extensive testing we applied our method to a producing
gold mine. This mine was operating on a schedule generated by MineMax. In order to
make a meaningful comparison with this schedule we simulated the same test
conditions used by MineMax. This involved reblocking the original block model which
contained 23 million blocks to one containing 1363 blocks. However, as reblocking
was carried out with different packages, the total value of the undiscounted pit used in
our model was 3.3% less. In this application T = 6 and the discount rate was 10%.
21
The constraints involved material movement and an upper bound on the milling
capacity (per period). Our software generated 7 good schedules within a total time of
10 minutes. Our best schedule was within 0.27% of the optimum and validated by
mining engineers as being realistic. Our schedule yielded an increase of 13.1% in the
NPV profit. In fact taking into account the differences in the block model our solution
value was at least 15% higher.
The MineMax solution (which was supplied to the mining company by the
software author) was obtained through a period by period optimization as the package
could not solve globally within the prescribed time limit. An important difference
between the two solutions is that ours generates a significantly higher cash flow in the
first two periods. This is in fact consistent with the aim of mine planners.
ACKNOWLEDGEMENT :
The authors gratefully acknowledge the Australian Research Council for
financial support through a SPIRT Grant (No. C69804881) and our industry partner
Optimum Planit for financial support and for assisting with all aspects of the project
including the provision of data.
REFERENCES :
[1] N.R. Achuthan, L. Caccetta and S.P. Hill, A New Subtour Elimination
Constraint for the Vehicle Routing Problem, European Journal of Operations
Research 91 (1995), pp. 573-586.
[2] N.R. Achuthan, L. Caccetta and S.P. Hill, Capacitated Vehicle Routing
Problem: Some New Cutting Planes, Asian-Pacific Journal of Operational
Research 15 (1998), pp. 109-123.
[3] R. Applegate, R. Bixby, V. Chvatal and W. Cook, Finding Cuts in the TSP (A
Preliminary Report), DIMACS Technical Report, (1995), pp. 95-105.
[4] P. Augerat, J.M. Belengeur, E. Benavent, A. Corberan, N. Naddef and G.
Rinaldi, Computational Results with a Branch and Cut Code for the
Capacitated Vehicle Routing Problem, Research Report 949-M, Universite
Joseph Fourier, Grenoble, France, (1995).
22
[5] Autoscheduler, Runge Mining Pty Ltd (Australia).
Web: http//www.runge.com/xpac.
[6] D.M. Francois-Bongarcon and D. Guibal, Parametization of Optimal Designs
of an Open Pit Beginning of a New Phase of Research, Trans. SME, AIME
274 (1984), 1801-1805.
[7] L. Caccetta and L.M. Giannini, Optimization Techniques for the Open Pit
Limit Problem, Proc. Australas. Inst. Min. Metall. 291 (1986), pp. 57-63.
[8] L. Caccetta and L.M. Giannini, An Application of Discrete mathematics in
the Design of an Open Pit Mine, Discrete Applied Mathematics 21 (1988), pp.
1-19.
[9] L. Caccetta and L.M. Giannini, Application of Operations Research
Techniques in Open Pit Mining, in Asian-Pacific Operations Research :
APORS88 (Byong-Hun Ahn Ed.), Elsevier Science Publishers BV (1990), pp.
707-724.
[10] L. Caccetta, L.M. Giannini and P. Kelsey, On the Implementation of Exact
Optimization Techniques for Open Pit Design, Asia-Pacific Journal of
Operations Research 11 (1994), pp. 155-170.
[11] L. Caccetta, L.M. Giannini and P. Kelsey, Application of Optimization
Techniques in Open Pit Mining, Proceedings of the Fourth International
Conference on Optimization Techniques and Applications (ICOTA98) (L.
Caccetta et al. Editors.), Vol. 1 (1998), pp. 414-422. (Curtin University of
Technology : Perth, Australia).
[12] L. Caccetta and S.P. Hill, Branch and Cut Methods for Network
Optimization, Mathematical and Computer Modelling (in press).
[13] L. Caccetta and S.P. Hill, A Branch and Cut Method for the Degree
Constrained Minimum Spanning Tree Problem, Networks (in press).
[14] L. Caccetta, P. Kelsey and L.M. Giannini, Open Pit Mine Production
Scheduling, in Computer Applications in the Minerals Industries International
Symposium (3
rd
Regional APCOM) (A.J. Basu, N. Stockton and D. Spottiswood
Editors), Austral. Inst. Min. Metall. Publication Series 5 (1998), pp. 65-72.
[15] T. Colou, Technical Parameterization for Open Pit Design and Mine
Planning, in : Proc. 21
st
APCOM Symposium of the Society of Mining
Engineers (AIME) (1988), pp. 485-494.
[16] CPLEX 6.0 User Manual, ILOG Inc., CPLEX Division, 889 Alder Avenue,
Suite 200, Incline Village, NV 89451, U.S.A.
[17] K. Dagdelen and T.B. Johnson, Optimum Open Pit Mine Production
Scheduling by Lagrangian Parameterization, in Proc. 19
th
APCOM
Symposium of the Society of Mining Engineers (AIME) (1986), pp. 127-142.
[18] B. Denby and D. Schofield, The Use of Genetic Algorithms in Underground
Mine Scheduling, in : Proc. 25
th
APCOM Symposium of the Society of Mining
Engineers (AIME, New York) (1995), pp. 389-394.
[19] B. Denby, D. Schofield and S. Bradford, Neural Network Applications in
Mining Engineering, Department of Mineral Resources Engineering Magazine,
University of Nottingham (1991), pp. 13-23.
[20] Earthworks, NPV Scheduler, Web: http://www.earthworks.com.au.
[21] M. Gershon, A Linear Programming Approach to Mine Scheduling
Optimization, in : Proc. 17
th
APCOM Symposium of the Society of Mining
Engineers (AIME) (1982), pp. 483-493.
[22] M. Gershon, Mine Scheduling Optimization with Mixed Integer
Programming, Mining Engineering 35(1983), pp. 351-354.
23
[23] M. Gershon, Heuristic Approaches for Mine Planning and Production
Scheduling, Int. Journal of Mining and Geological Engineering 5 (1987), pp. 1-
13.
[24] L. Gignac, Computerized Ore Evaluation and Open Pit Design, in : Proc.
36
th
Annual Mining Symposium of the Society of Mining Engineers (AIME)
(1975), pp. 45-53.
[25] K.L. Hoffman and M.W. Padberg, Solving Airline Crew Scheduling
Problems by Branch and Cut, Management Science 39 (1993), pp. 657-682.
[26] Y.C. Kim, Open-Pit Limits Analysis : Technical Overview, in : Computer
Methods for the 80s (A. Weiss, Editor), (AIME, New York) (1979), pp. 297-
303.
[27] Y.C. Kim, Production Scheduling : Technical Overview, in : Computer
Methods for the 80s (A. Weiss, Editor) (AIME, New York) (1979), pp. 610-
614.
[28] H. Lerchs and I.F. Grossmann, Optimum Design of Open Pit Mines, Canad.
Inst. Mining Bull. 58 (1965), pp. 47-54.
[29] G. Matheron, Le Parametrage des Contours Optimaux, Technical Report 403
(1975), Centre de Geostatistiques, Fontainebleau, France.
[30] G. Matheron, Le Parametrage Technique des Reseues, Technical Report No.
453 (1975), Centre de Geostatistiques, Fontainebleau, France.
[31] MineMax, Combinatorics Pty Ltd, Unit 4b, R&D Centre, 1 Sarich Way,
Bentley, Western Australia, 6102.
[32] M. Padberg and G. Rinaldi, A Branch and Cut Algorithm for the Resolution
of Large Scale Travelling Salesman Problems, SIAM Review 33 No. 1
(1991), pp. 60-100.
[33] J.C. Picard, Maximum Closure of a Graph and Applications to
Combinatorial Problems, Management Sc. 22 (1976), pp. 1268-1272.
[34] G. Thomas, Optimization of Mine Production Scheduling - the State of the
Art, in Proceedings IIR Dollar Driven Planning Conference (1996).
[35] G. Thomas, Pit Optimization and Mine Production Scheduling - The Way
Ahead, in : Proceed. 26
th
APCOM Symposium of the Society of Mining
Engineers (AIME) (1976), pp. 221-228.
[36] B. Tolwinski and R. Underwood, A Scheduling Algorithm for Open Pit
Mines, IMA Journal of Mathematics Applied in Business and Industry 7 (1996),
pp. 247-270.
[37] J. Whittle, Four-D User Manual (1993). Whittle Programming Pty Ltd.,
Melbourne, Australia.
[38] J Whittle, Four-X User Manual (1998). Whittle Programming Pty Ltd.,
Melbourne, Australia.
[39] J. Whittle, Open Pit Optimization, Surface Mining (2
nd
Edition), AMIE
(1990), pp. 470-475.
[40] J. Yamatomi, G. Mogi, A. Akaike and U. Yamaguchi, Selection Extraction
Dynamic Cone Algorithm for Three-Dimensional Open Pit Designs, in :
Proceed. 25
th
APCOM Symposium of the Society of Mining Engineers (AIME)
(1995), pp. 267-274.
paper19.doc
24